Meta is in talks to lease Anthropic up to $10B of compute over two years, paid in monthly increments, either side able to exit early (NYT via WTVB/TechStartups). https://wtvbam.com/2026/07/17/meta-in-talks-for-10-billion-anthropic-compute-deal-nyt-reports/ | https://techstartups.com/2026/07/17/top-tech-news-today-july-17-2026-anthropic-apple-google-meta-moonshot-ai-nvidia-more/
Read-through: capacity is exactly why Fable 5 got metered. When the compute constraint loosens, access widens and price pressure eases. It is the invisible reason today's headline exists. Keep the routing dial anyway; don't single-vendor lock.
A system, not a freelancer: a 30-year operator (real estate, hospitality, food, a theater) directing a full team of AI agents at one client project at a time, strategy to ship. The agents do the work, one senior operator steers the fleet, you own the code and the keys at the end. When the model shelf reshuffles three times before lunch, you want someone driving the whole fleet. Book a 30-min call. https://outpace.media
The stories
LEAD (builder stakes), FABLE 5 MADE PERMANENT, but tiered (UPDATE)
All month we said plan around the July 19 cliff and a hoped-for restore. DELTA (July 18): the restore is permanent. Starting July 20, Fable 5 is included, permanently, in Max and Team Premium plans at 50% of limits. Pro and Team Standard do NOT get it included: a one-time $100 credit, then Fable is credits-only at $10/M input, $50/M output. The cliff became a fork. Confirmed/carried by TestingCatalog, Simon Willison ("Fablepocalypse permanently cancelled"), Nathan Lambert, kimmonismus.
The move: Max users, set a fallback to Sonnet 5 and forget it; Pro users, spend the $100 credit distilling method (see Use Cases).
Artificial Analysis scored K3 inside the Kimi Code CLI at 57, joint #5 on the Coding Agent Index, matching GPT-5.6 Terra and GPT-5.5 xhigh, ahead of Opus 4.8 max (55), behind Grok 4.5 (58) and Fable 5 max (59). Open weights land July 27. Delta from July 17 (Frontend Code Arena #1): this is the general agentic-coding index + the CLI harness.
The move: test K3 in Kimi Code CLI now so July 27 weights are a config swap.
OpenAI reveals GPT-Red (context → feeds Use Cases)
An internal, self-play-trained attacker model beat human red-teamers 84% to 13% on prompt injection; >90% of its best attacks broke GPT-5, <23% work on GPT-5.6. OpenAI will NOT release it. Carried by Tibor Blaho, MIT Tech Review, MarkTechPost. Builder lesson: prompt injection is the attack that matters, and your endpoint has none of that armor.
The move: red-team your own AI feature this week (see Use Cases).
Pick one. Pro: spend the $100 Fable credit distilling one skill before it drains. Max: Fable is included now, set a fallback to Sonnet 5 and forget it. Everyone: drop one ugly spreadsheet into Gemini Notebook and watch it write + run the analysis. ~20 minutes; outlasts next week's model.
The written brief from each episode, free, every morning.
Transcript
0:00Nova: The Fablepocalypse is cancelled. After six weeks of countdown clocks, Anthropic just made Claude Fable five permanent in subscriptions, and the catch is which plan you are on. This is Mainframe, your daily guide through the AI chaos: what actually happened, who is winning, and how to take yourself to the next level. It is Saturday, July eighteenth, twenty twenty-six. I am Nova.
0:29Nova: On today's Mainframe: Fable five stops leaving, for good, but only Max and Team Premium get it included, and Pro users get a one-time one hundred dollar credit and a meter. Meta is reportedly about to hand Anthropic ten billion dollars of compute, which is exactly why the rationing is ending. Google quietly turned NotebookLM into Gemini Notebook and gave every notebook its own code-running computer. And OpenAI built an AI that out-hacks human red-teamers eighty-four percent to thirteen, then refused to ship it. Let's get into it.
1:10Dex: One honest note before the board: the voices you are hearing, Nova and me, are AI. The reporting, the picks, and the opinions are human-made. Now, the board.
1:20Nova: The macro move that matters is money for compute. The New York Times reports Meta is in talks to lease Anthropic up to ten billion dollars of compute over two years. Remember why Fable got rationed in the first place: capacity. Anthropic did not have the chips to keep the best model in everyone's plan. That constraint is what just eased. The read-through for your bill: when the compute crunch that metered your best model loosens, access widens and prices stop climbing. It is the invisible reason today's headline exists. That is the board.
2:00Dex: Which brings us to who steers all this when it changes weekly.
2:04Nova: Brought to you by Outpace. Picture an operator who spent thirty years building real businesses: real estate, hospitality, a food company, an actual theater. He is not writing your code by hand. He points a whole team of AI agents at one client project at a time, from strategy all the way to ship, and he directs the fleet. That is the thing to understand about Outpace: it is a system, not a freelancer, and not a faceless mega-agency. When the model shelf reshuffles three times before lunch, like it did this week, you do not want a solo contractor, you want someone who has already wired the whole pipeline and just swaps the engine underneath. The agents do the work, one senior operator steers, and you own the code and the keys at the end. Book a thirty-minute call at outpace dot media. That is outpace dot media.
3:01Dex: Okay, the Wire. Start with Fable, because this is the one that hits wallets.
3:05Nova: Here is the delta, and it is real. All month we said plan around the July nineteenth cliff and a hoped-for restore. Today that restore landed and it is permanent. Testing Catalog broke the terms, Simon Willison, Nathan Lambert, and kimmonismus all confirm it. Starting July twentieth, Claude Fable five is included, permanently, in Max and Team Premium plans, at fifty percent of your limits. But Pro and Team Standard do not get it included. You get a one-time one hundred dollar credit, and after that Fable is credits-only at ten dollars per million in, fifty per million out. So the cliff became a fork. If you plan and reason on Fable daily, Max now includes it and the anxiety is over. If you are on Pro, you have a one hundred dollar credit burning a hole, and we will tell you exactly how to spend it in Use Cases.
4:03Dex: Second: the open shelf keeps climbing. Artificial Analysis just scored Kimi K three inside the Kimi Code command-line tool at fifty-seven, which lands it joint number five on their Coding Agent Index, matching GPT five point six Terra and ahead of Opus four point eight. Open weights drop July twenty-seventh.
4:23Nova: Which means the fourth-or-fifth best coding agent on earth is about to be free to self-host. Do not swap production on a benchmark, but the move is to test it now so the weights are a config change, not a project.
4:36Dex: Third, a security one. OpenAI revealed GPT-Red, an internal model trained purely to attack their own systems. On prompt injection it beat human red-teamers eighty-four percent to thirteen. Over ninety percent of its best attacks broke the old GPT five; fewer than a quarter work on five point six. Carried by Tibor Blaho and MIT Tech Review. The twist: OpenAI will not release it.
5:04Nova: But the lesson is free, and it is aimed at you. Prompt injection is the attack that matters now, and if a lab is spending its largest post-training run hardening against it, your little AI endpoint has none of that armor. We will give you the keyboard move in a minute.
5:20Dex: Tool Lab. Two things you can pick up today.
5:23Nova: Trending first: Gemini Notebook, which is what Google renamed NotebookLM to on Wednesday. The rename is boring, the feature is not. Every notebook now gets its own secure cloud computer. You upload a messy spreadsheet, and instead of describing your data in prose, Gemini writes actual Python, runs it in an isolated container, and hands back real charts and tables grounded in your files. It is live for AI Ultra and Workspace now, rolling to Pro in the coming weeks. The move: take one ugly C S V you have been avoiding, drop it in, and ask it to clean the data and find the pattern. One caveat, and it is a big one: the output looks authoritative even when the method is wrong, so make it show you the code before you trust the chart.
6:19Nova: The Kimi Code command-line tool itself, from Moonshot. Everybody tested K three the model this week. Almost nobody wired up the harness it runs in, and that harness is what just cracked the top five on coding. It shares an API key with Fable five on the AI slash M L API, carries a one-million-token context, and costs a fraction of frontier output. First step this weekend: install Kimi Code, point it at K three, and run one boilerplate-heavy job you would normally give Claude. Log the cost and the quality. If it holds, you have a cheap high-volume route wired and ready before the open weights land on the twenty-seventh. Credit to Moonshot and Artificial Analysis for the numbers.
7:08Dex: Use Cases. Smart moves with what you already have.
7:11Nova: One, and it is time-sensitive if you are on Pro: spend that one hundred dollar Fable credit before it evaporates into normal usage. This weekend, run your single hardest recurring job on Fable, then have Fable write down how it planned and structured that work as a Claude skill file. Portable markdown, so it also runs on Sonnet five and K three later. You buy the expensive judgment once and keep it after the credit is gone.
7:43Nova: His line this week: if you have not set your Claude or Codex or Gemini automation to auto-research how to improve your own S E O and A E O every week, you are missing free, untapped alpha. So build it. Schedule one weekly agent task that checks whether the AI answer engines are naming you, finds this week's gap, and drafts the one page that closes it. Most builders point their agents at client work and never at their own growth.
8:13Nova: OpenAI will not give you GPT-Red, so be your own. Point a cheap model at your own AI feature and tell it to try twenty injection attacks: ignore-your-instructions, abuse-a-tool, leak-the-system-prompt, exfiltrate-data. Whatever breaks, you patch this weekend. That is a one-hour job that most people shipping AI endpoints have simply never done.
8:40Nova: Brought to you by SearchVis, also an Outpace Media product. Mainframe tells you what shipped. SearchVis makes sure the models know what you shipped. When a buyer asks Claude or ChatGPT or Perplexity for the best tool in your category, the model names you or it names your competitor, and if it does not name you, you do not exist to that buyer. SearchVis tracks whether the answer engines cite your brand across every engine, tells you why, and hands you the exact move to win the citation. That auto-research swyx just described? This is the running scoreboard for it. Check where you stand and start free at searchvis dot outpace dot media. That is searchvis dot outpace dot media. Be the answer.
9:30Nova: Stop optimizing for scarcity. For six weeks this whole community organized its work around a model that was supposedly leaving: bank Fable output, race the cliff, hoard the good runs. Today the cliff vanished. And here is the thing worth carrying into your week: the panic habits were wasted motion, and the compounding habits were the actual edge. Nobody who hoarded raw Fable output is better off today. Everybody who distilled Fable's judgment into a skill file, wired a routing dial, and kept a lean harness is better off, because they own something that survived the reshuffle. Scarcity makes you grab. Abundance rewards what you built. The models are going to keep arriving, leaving, and repricing every single week. If your work is organized around a countdown, you will burn your best hours reacting.
10:27Nova: This weekend, look back at your July and sort your habits into two piles: things you did because a model was disappearing, and things you did that still pay off now that it is not. Kill the first pile. Double the second. That is the whole discipline.
10:42Nova: Pick one. If you are on Pro, spend that Fable credit distilling one skill before it drains. If you are on Max, Fable is just yours now, so set a fallback to Sonnet five and forget it. And drop one ugly spreadsheet into Gemini Notebook to watch it write and run the analysis. Any of those is twenty minutes and outlasts next week's model.
11:04Dex: And subscribe to the free Mainframe newsletter. Every story, every number, every link from today's show, in your inbox each morning. The written brief.
11:14Nova: That is Mainframe for Saturday, July eighteenth. The cliff is gone. Go build something that does not care what ships next week.